Unverified85% confidenceFactTime unknown
OmniMerge V4在MBPP代码测试中达到73.4%，相比原版Q6K量化的56.2%提升了15.8个百分点
1
Sources
85%
Confidence
Medium-term (~90 days)
Relevance
6/1/2026
First Seen
Valid until: 8/30/2026
Sources
Qwen3.6 27B三大邪修量化模型实测：代码暴增15.8PP、40B蒸馏、16GB适配
bilibili尚书省说书人
Related Entities
Related Claims
UnverifiedOmniMerge V4在HumanEval测试中得分84.76，与Q6K原版持平71% similarUnverifiedDeepSeek V4 Pro在CodeForces取得3206分，MMT-266评测达到95.2%通过率67% similarUnverifiedQwen3.5的参数激活率为4.3%，而DeepSeek V3的671B模型激活率为5.5%67% similarUnverifiedGLM4在代码推理和通用知识测试中得分74.4%，仅次于Claude 3 Opus66% similarUnverifiedMistral 3.8B 指令模型使用 Q4 量化和 llama.cpp 原生调用时，综合 Score 从 27.3% 提升到 78.3%，区别仅在于是否有 Forge 护栏66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/18009API
curl https://kongchang.com/api/v1/knowledge/claims/18009MCP
get_claim(id=18009)